# Splink Record Linkage

> Open-source Python library from the UK Ministry of Justice for probabilistic record linkage and deduplication, based on the Fellegi-Sunter model, running on DuckDB, Spark, PostgreSQL, or SQLite without training data.

Source: https://toolsfordata.com/tools/splink/ · Updated 2026-11-19

[Website](https://moj-analytical-services.github.io/splink/) · [Docs](https://moj-analytical-services.github.io/splink/getting_started.html) · [GitHub](https://github.com/moj-analytical-services/splink) (2.6k stars)

Best for: data teams that need free, explainable entity resolution in Python, from a million records on a laptop to 100+ million on Spark, and build the MDM hub themselves.

## Overview

- Core Strength: Free probabilistic record linkage in Python, on DuckDB or Spark
- Matching: Fellegi-Sunter model estimated without labels
- Golden Records: None: outputs clusters, golden records are up to you
- Platform Fit: DuckDB, Spark, PostgreSQL, SQLite
- Deployment Model: Self-hosted Python library
- Pricing: Free (MIT)
- Notable Caveat: No golden records, no persistent IDs, no stewardship

## What it's used for

Splink is used to deduplicate and link records in Python without training data: it estimates probabilistic match weights and runs on DuckDB for a million records on a laptop, or on Spark for 100+ million.

## Pricing

Splink is free under the MIT license, with no paid edition.

## Features

### Matching

- ML Matching: Yes — Fellegi-Sunter probabilistic model with weights estimated by EM: "No training data is required for model training." ([source](https://moj-analytical-services.github.io/splink/))
- Rule-Based Matching: Yes — Comparisons with exact and fuzzy levels, and a deterministic_link mode with strict rules. ([source](https://moj-analytical-services.github.io/splink/api_docs/inference.html))
- Real-Time Matching: No — compare_two_records scores a pair you pass in a Python call; there is no service that matches a record against existing entities. ([source](https://moj-analytical-services.github.io/splink/demos/examples/duckdb/real_time_record_linkage.html))
- Incremental Matching: No — predict_between finds links between existing and new records; updates and deletes are not handled. ([source](https://moj-analytical-services.github.io/splink/api_docs/inference.html))
- Persistent Entity ID: No — Clusters are connected components, with ties resolved by "lowest_id"; IDs are not stable across runs. ([source](https://moj-analytical-services.github.io/splink/api_docs/linker_clustering.html))
- Match Explainability: Yes — The waterfall chart "shows the amount of evidence of a match that is provided by each comparison". ([source](https://moj-analytical-services.github.io/splink/charts/waterfall_chart.html))

### Golden Records

- Survivorship Rules: No — Splink outputs clusters of IDs; building golden records is left to you. ([source](https://moj-analytical-services.github.io/splink/api_docs/linker_clustering.html))
- Survivorship on Read: No — Splink outputs clusters of IDs; no golden records. ([source](https://moj-analytical-services.github.io/splink/api_docs/linker_clustering.html))
- Stewardship Review: No — The cluster studio dashboard visualizes clusters but doesn't approve or split production matches. ([source](https://moj-analytical-services.github.io/splink/charts/cluster_studio_dashboard.html))
- Record History: No — No record versioning or merge history. ([source](https://moj-analytical-services.github.io/splink/))
- Hierarchy Management: No — No master-record hierarchies. ([source](https://moj-analytical-services.github.io/splink/))

### Domains & Modeling

- Multi-Domain MDM: No — A library that links the tables you pass it; no managed domains. ([source](https://moj-analytical-services.github.io/splink/))
- Configurable Data Model: No — A settings object defines comparisons, not a stored master data model. ([source](https://moj-analytical-services.github.io/splink/topic_guides/splink_fundamentals/settings.html))
- Product Information Management: No — No PIM features. ([source](https://moj-analytical-services.github.io/splink/))
- Reference Data Management: No — No code-list management. ([source](https://moj-analytical-services.github.io/splink/))
- Master Data Authoring: No — A Python library with no entry UI. ([source](https://moj-analytical-services.github.io/splink/))

### Data Quality

- Validation at Entry: No — No entry UI or API. ([source](https://moj-analytical-services.github.io/splink/))
- Standardization: No — "Before using Splink, input data should be standardised": cleaning is left to you. ([source](https://moj-analytical-services.github.io/splink/))
- Third-Party Enrichment: No — No enrichment connectors. ([source](https://moj-analytical-services.github.io/splink/))

### Data Distribution

- REST API: No — "Splink is a Python package". ([source](https://github.com/moj-analytical-services/splink))
- Change Events: No — No event publishing. ([source](https://moj-analytical-services.github.io/splink/topic_guides/splink_fundamentals/backends/backends.html))
- Warehouse Write-Back: No — Results live in the backend you run (DuckDB, Spark, SQLite, PostgreSQL); no warehouse write-back feature. ([source](https://moj-analytical-services.github.io/splink/topic_guides/splink_fundamentals/backends/backends.html))
- Data Syndication: No — No publishing to external channels. ([source](https://moj-analytical-services.github.io/splink/))

### AI & Agents

- MCP Server: No — No MCP server. ([source](https://moj-analytical-services.github.io/splink/))
- AI Stewardship Assistance: No — No AI assistant for steward decisions. ([source](https://moj-analytical-services.github.io/splink/))

### Platform Native

- In-Place Matching: Yes — Splink generates SQL and submits it "to a backend of the user's choosing for execution": DuckDB, Spark, PostgreSQL or SQLite. ([source](https://moj-analytical-services.github.io/splink/topic_guides/splink_fundamentals/backends/backends.html))
- Snowflake Native: No — Snowflake is not a supported backend. ([source](https://moj-analytical-services.github.io/splink/topic_guides/splink_fundamentals/backends/backends.html))
- Databricks Native: Unknown — Runs on Spark, but "the Splink development team does not have access to a Databricks environment", so Databricks isn't a documented target. ([source](https://moj-analytical-services.github.io/splink/topic_guides/splink_fundamentals/backends/backends.html))
- Microsoft Fabric Native: No — Fabric is not a documented backend. ([source](https://moj-analytical-services.github.io/splink/topic_guides/splink_fundamentals/backends/backends.html))

### Deployment

- SaaS Deployment: No — A Python package ("pip install splink"), with no hosted service. ([source](https://github.com/moj-analytical-services/splink))
- Self-Hosted Deployment: Yes — "It can be run on any device with python installed". ([source](https://moj-analytical-services.github.io/splink/topic_guides/splink_fundamentals/backends/backends.html))

### Open Source

- Open Source: Yes — MIT license. ([source](https://github.com/moj-analytical-services/splink))

## Integrations

4 integrations, each sourced from the vendor docs.

- Data Platform: Apache Spark
- Databases: DuckDB, SQLite, PostgreSQL

## Alternatives

- [Zingg](https://toolsfordata.com/tools/zingg/): The closest open-source match: ML matching trained from labeled pairs, with persistent entity IDs and incremental runs in the paid editions.
- [Senzing](https://toolsfordata.com/tools/senzing/): Consider Senzing for real-time entity resolution embedded in your own application, with no model to train.
- [AWS Entity Resolution](https://toolsfordata.com/tools/aws-entity-resolution/): Consider AWS Entity Resolution for managed rule-based or ML matching on S3 data, priced per record.
